2 * Copyright 2017 Omnibond Systems, L.L.C.
6 #include "orangefs-kernel.h"
7 #include "orangefs-bufmap.h"
9 struct orangefs_dir_part {
10 struct orangefs_dir_part *next;
16 struct orangefs_dir_part *part;
21 #define PART_SHIFT (24)
22 #define PART_SIZE (1<<24)
23 #define PART_MASK (~(PART_SIZE - 1))
26 * There can be up to 512 directory entries. Each entry is encoded as
28 * 4 bytes: string size (n)
30 * 1 byte: trailing zero
35 * The trailer_buf starts with a struct orangefs_readdir_response_s
36 * which must be skipped to get to the directory data.
38 * The data which is received from the userspace daemon is termed a
39 * part and is stored in a linked list in case more than one part is
40 * needed for a large directory.
42 * The position pointer (ctx->pos) encodes the part and offset on which
43 * to begin reading at. Bits above PART_SHIFT encode the part and bits
44 * below PART_SHIFT encode the offset. Parts are stored in a linked
45 * list which grows as data is received from the server. The overhead
46 * associated with managing the list is presumed to be small compared to
47 * the overhead of communicating with the server.
49 * As data is received from the server, it is placed at the end of the
50 * part list. Data is parsed from the current position as it is needed.
51 * When data is determined to be corrupt, it is either because the
52 * userspace component has sent back corrupt data or because the file
53 * pointer has been moved to an invalid location. Since the two cannot
54 * be differentiated, return EIO.
56 * Part zero is synthesized to contains `.' and `..'. Part one is the
57 * first part of the part list.
60 static int do_readdir(struct orangefs_inode_s *oi,
61 struct orangefs_dir *od, struct dentry *dentry,
62 struct orangefs_kernel_op_s *op)
64 struct orangefs_readdir_response_s *resp;
68 * Despite the badly named field, readdir does not use shared
69 * memory. However, there are a limited number of readdir
70 * slots, which must be allocated here. This flag simply tells
71 * the op scheduler to return the op here for retry.
73 op->uses_shared_memory = 1;
74 op->upcall.req.readdir.refn = oi->refn;
75 op->upcall.req.readdir.token = od->token;
76 op->upcall.req.readdir.max_dirent_count =
77 ORANGEFS_MAX_DIRENT_COUNT_READDIR;
80 bufi = orangefs_readdir_index_get();
86 op->upcall.req.readdir.buf_index = bufi;
88 r = service_operation(op, "orangefs_readdir",
89 get_interruptible_flag(dentry->d_inode));
91 orangefs_readdir_index_put(bufi);
93 if (op_state_purged(op)) {
95 vfree(op->downcall.trailer_buf);
97 } else if (r == -EIO) {
98 vfree(op->downcall.trailer_buf);
105 vfree(op->downcall.trailer_buf);
108 } else if (op->downcall.status) {
109 vfree(op->downcall.trailer_buf);
110 od->error = op->downcall.status;
111 return op->downcall.status;
115 * The maximum size is size per entry times the 512 entries plus
116 * the header. This is well under the limit.
118 if (op->downcall.trailer_size > PART_SIZE) {
119 vfree(op->downcall.trailer_buf);
124 resp = (struct orangefs_readdir_response_s *)
125 op->downcall.trailer_buf;
126 od->token = resp->token;
130 static int parse_readdir(struct orangefs_dir *od,
131 struct orangefs_kernel_op_s *op)
133 struct orangefs_dir_part *part, *new;
146 new = (void *)op->downcall.trailer_buf;
148 new->len = op->downcall.trailer_size -
149 sizeof(struct orangefs_readdir_response_s);
155 od->end = count << PART_SHIFT;
160 static int orangefs_dir_more(struct orangefs_inode_s *oi,
161 struct orangefs_dir *od, struct dentry *dentry)
163 struct orangefs_kernel_op_s *op;
166 op = op_alloc(ORANGEFS_VFS_OP_READDIR);
171 r = do_readdir(oi, od, dentry, op);
176 r = parse_readdir(od, op);
188 static int fill_from_part(struct orangefs_dir_part *part,
189 struct dir_context *ctx)
191 const int offset = sizeof(struct orangefs_readdir_response_s);
192 struct orangefs_khandle *khandle;
196 i = ctx->pos & ~PART_MASK;
198 /* The file offset from userspace is too large. */
203 * If the seek pointer is positioned just before an entry it
204 * should find the next entry.
209 while (i < part->len) {
210 if (part->len < i + sizeof *len)
212 len = (void *)part + offset + i;
214 * len is the size of the string itself. padlen is the
215 * total size of the encoded string.
217 padlen = (sizeof *len + *len + 1) +
218 (8 - (sizeof *len + *len + 1)%8)%8;
219 if (part->len < i + padlen + sizeof *khandle)
221 s = (void *)part + offset + i + sizeof *len;
224 khandle = (void *)part + offset + i + padlen;
225 if (!dir_emit(ctx, s, *len,
226 orangefs_khandle_to_ino(khandle),
229 i += padlen + sizeof *khandle;
231 BUG_ON(i > part->len);
232 ctx->pos = (ctx->pos & PART_MASK) | i;
240 static int orangefs_dir_fill(struct orangefs_inode_s *oi,
241 struct orangefs_dir *od, struct dentry *dentry,
242 struct dir_context *ctx)
244 struct orangefs_dir_part *part;
247 count = ((ctx->pos & PART_MASK) >> PART_SHIFT) - 1;
250 while (part->next && count) {
254 /* This means the userspace file offset is invalid. */
260 while (part && part->len) {
262 r = fill_from_part(part, ctx);
267 /* Userspace buffer is full. */
271 * The part ran out of data. Move to the next
273 ctx->pos = (ctx->pos & PART_MASK) +
281 static loff_t orangefs_dir_llseek(struct file *file, loff_t offset,
284 struct orangefs_dir *od = file->private_data;
286 * Delete the stored data so userspace sees new directory
289 if (!whence && offset < od->end) {
290 struct orangefs_dir_part *part = od->part;
292 struct orangefs_dir_part *next = part->next;
296 od->token = ORANGEFS_ITERATE_START;
298 od->end = 1 << PART_SHIFT;
300 return default_llseek(file, offset, whence);
303 static int orangefs_dir_iterate(struct file *file,
304 struct dir_context *ctx)
306 struct orangefs_inode_s *oi;
307 struct orangefs_dir *od;
308 struct dentry *dentry;
311 dentry = file->f_path.dentry;
312 oi = ORANGEFS_I(dentry->d_inode);
313 od = file->private_data;
319 if (!dir_emit_dot(file, ctx))
324 if (!dir_emit_dotdot(file, ctx))
326 ctx->pos = 1 << PART_SHIFT;
330 * The seek position is in the first synthesized part but is not
333 if ((ctx->pos & PART_MASK) == 0)
339 * Must read more if the user has sought past what has been read
340 * so far. Stop a user who has sought past the end.
342 while (od->token != ORANGEFS_ITERATE_END &&
343 ctx->pos > od->end) {
344 r = orangefs_dir_more(oi, od, dentry);
348 if (od->token == ORANGEFS_ITERATE_END && ctx->pos > od->end)
351 /* Then try to fill if there's any left in the buffer. */
352 if (ctx->pos < od->end) {
353 r = orangefs_dir_fill(oi, od, dentry, ctx);
358 /* Finally get some more and try to fill. */
359 if (od->token != ORANGEFS_ITERATE_END) {
360 r = orangefs_dir_more(oi, od, dentry);
363 r = orangefs_dir_fill(oi, od, dentry, ctx);
369 static int orangefs_dir_open(struct inode *inode, struct file *file)
371 struct orangefs_dir *od;
372 file->private_data = kmalloc(sizeof(struct orangefs_dir),
374 if (!file->private_data)
376 od = file->private_data;
377 od->token = ORANGEFS_ITERATE_START;
379 od->end = 1 << PART_SHIFT;
384 static int orangefs_dir_release(struct inode *inode, struct file *file)
386 struct orangefs_dir *od = file->private_data;
387 struct orangefs_dir_part *part = od->part;
388 orangefs_flush_inode(inode);
390 struct orangefs_dir_part *next = part->next;
398 const struct file_operations orangefs_dir_operations = {
399 .llseek = orangefs_dir_llseek,
400 .read = generic_read_dir,
401 .iterate = orangefs_dir_iterate,
402 .open = orangefs_dir_open,
403 .release = orangefs_dir_release